Attention

The vector search and clustering algorithms in RAFT are being migrated to a new library dedicated to vector search called cuVS. We will continue to support the vector search algorithms in RAFT during this move, but will no longer update them after the RAPIDS 24.06 (June) release. We plan to complete the migration by RAPIDS 24.08 (August) release.

IVF-PQ#

#include <raft/neighbors/ivf_pq.cuh>

namespace raft::neighbors::ivf_pq

enum class codebook_gen#

A type for specifying how PQ codebooks are created.

Values:

enumerator PER_SUBSPACE#
enumerator PER_CLUSTER#
template<typename IdxT, typename SizeT = uint32_t>
using list_data = ivf::list<list_spec, SizeT, IdxT>#
static constexpr uint32_t kIndexGroupSize = 32

Size of the interleaved group.

static constexpr uint32_t kIndexGroupVecLen = 16#

Stride of the interleaved group for vectorized loads.

template<typename IdxT>
static constexpr IdxT kOutOfBoundsRecord = std::numeric_limits<IdxT>::max()#

Default value returned by search when the n_probes is too small and top-k is too large. One may encounter it if the combined size of probed clusters is smaller than the requested number of results per query.

template<typename T, typename IdxT = uint32_t>
index<IdxT> build(raft::resources const &handle, const index_params &params, raft::device_matrix_view<const T, IdxT, row_major> dataset)#

Build the index from the dataset for efficient search.

NB: Currently, the following distance metrics are supported:

  • L2Expanded

  • L2Unexpanded

  • InnerProduct

Template Parameters:
  • T – data element type

  • IdxT – type of the indices in the source dataset

Parameters:
  • handle[in]

  • params[in] configure the index building

  • dataset[in] a device matrix view to a row-major matrix [n_rows, dim]

Returns:

the constructed ivf-pq index

template<typename T, typename IdxT>
index<IdxT> extend(raft::resources const &handle, raft::device_matrix_view<const T, IdxT, row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT>> new_indices, const index<IdxT> &idx)#

Extend the index with the new data. *.

Template Parameters:
  • T – data element type

  • IdxT – type of the indices in the source dataset

Parameters:
  • handle[in]

  • new_vectors[in] a device matrix view to a row-major matrix [n_rows, idx.dim()]

  • new_indices[in] a device vector view to a vector of indices [n_rows]. If the original index is empty (idx.size() == 0), you can pass std::nullopt here to imply a continuous range [0...n_rows).

  • idx[inout]

template<typename T, typename IdxT>
void extend(raft::resources const &handle, raft::device_matrix_view<const T, IdxT, row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT>> new_indices, index<IdxT> *idx)#

Extend the index with the new data. *.

Template Parameters:
  • T – data element type

  • IdxT – type of the indices in the source dataset

Parameters:
  • handle[in]

  • new_vectors[in] a device matrix view to a row-major matrix [n_rows, idx.dim()]

  • new_indices[in] a device vector view to a vector of indices [n_rows]. If the original index is empty (idx.size() == 0), you can pass std::nullopt here to imply a continuous range [0...n_rows).

  • idx[inout]

template<typename T, typename IdxT, typename IvfSampleFilterT>
void search_with_filtering(raft::resources const &handle, const search_params &params, const index<IdxT> &idx, raft::device_matrix_view<const T, uint32_t, row_major> queries, raft::device_matrix_view<IdxT, uint32_t, row_major> neighbors, raft::device_matrix_view<float, uint32_t, row_major> distances, IvfSampleFilterT sample_filter = IvfSampleFilterT{})#

Search ANN using the constructed index with the given filter.

See the ivf_pq::build documentation for a usage example.

Note, this function requires a temporary buffer to store intermediate results between cuda kernel calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can pass a pool memory resource or a large enough pre-allocated memory resource to reduce or eliminate entirely allocations happening within search. The exact size of the temporary buffer depends on multiple factors and is an implementation detail. However, you can safely specify a small initial size for the memory pool, so that only a few allocations happen to grow it during the first invocations of the search.

Template Parameters:
  • T – data element type

  • IdxT – type of the indices

  • IvfSampleFilterT – Device filter function, with the signature (uint32_t query_ix, uint32 cluster_ix, uint32_t sample_ix) -> bool or (uint32_t query_ix, uint32 sample_ix) -> bool

Parameters:
  • handle[in]

  • params[in] configure the search

  • idx[in] ivf-pq constructed index

  • queries[in] a device matrix view to a row-major matrix [n_queries, index->dim()]

  • neighbors[out] a device matrix view to the indices of the neighbors in the source dataset [n_queries, k]

  • distances[out] a device matrix view to the distances to the selected neighbors [n_queries, k]

  • sample_filter[in] a device filter function that greenlights samples for a given query.

template<typename T, typename IdxT>
void search(raft::resources const &handle, const search_params &params, const index<IdxT> &idx, raft::device_matrix_view<const T, uint32_t, row_major> queries, raft::device_matrix_view<IdxT, uint32_t, row_major> neighbors, raft::device_matrix_view<float, uint32_t, row_major> distances)#

Search ANN using the constructed index.

See the ivf_pq::build documentation for a usage example.

Note, this function requires a temporary buffer to store intermediate results between cuda kernel calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can pass a pool memory resource or a large enough pre-allocated memory resource to reduce or eliminate entirely allocations happening within search. The exact size of the temporary buffer depends on multiple factors and is an implementation detail. However, you can safely specify a small initial size for the memory pool, so that only a few allocations happen to grow it during the first invocations of the search.

Template Parameters:
  • T – data element type

  • IdxT – type of the indices

Parameters:
  • handle[in]

  • params[in] configure the search

  • idx[in] ivf-pq constructed index

  • queries[in] a device matrix view to a row-major matrix [n_queries, index->dim()]

  • neighbors[out] a device matrix view to the indices of the neighbors in the source dataset [n_queries, k]

  • distances[out] a device matrix view to the distances to the selected neighbors [n_queries, k]

struct index_params : public raft::neighbors::ann::index_params#
#include <ivf_pq_types.hpp>

Public Members

uint32_t n_lists = 1024#

The number of inverted lists (clusters)

Hint: the number of vectors per cluster (n_rows/n_lists) should be approximately 1,000 to 10,000.

uint32_t kmeans_n_iters = 20#

The number of iterations searching for kmeans centers (index building).

double kmeans_trainset_fraction = 0.5#

The fraction of data to use during iterative kmeans building.

uint32_t pq_bits = 8#

The bit length of the vector element after compression by PQ.

Possible values: [4, 5, 6, 7, 8].

Hint: the smaller the ‘pq_bits’, the smaller the index size and the better the search performance, but the lower the recall.

uint32_t pq_dim = 0#

The dimensionality of the vector after compression by PQ. When zero, an optimal value is selected using a heuristic.

NB: pq_dim * pq_bits must be a multiple of 8.

Hint: a smaller ‘pq_dim’ results in a smaller index size and better search performance, but lower recall. If ‘pq_bits’ is 8, ‘pq_dim’ can be set to any number, but multiple of 8 are desirable for good performance. If ‘pq_bits’ is not 8, ‘pq_dim’ should be a multiple of 8. For good performance, it is desirable that ‘pq_dim’ is a multiple of 32. Ideally, ‘pq_dim’ should be also a divisor of the dataset dim.

codebook_gen codebook_kind = codebook_gen::PER_SUBSPACE#

How PQ codebooks are created.

bool force_random_rotation = false#

Apply a random rotation matrix on the input data and queries even if dim % pq_dim == 0.

Note: if dim is not multiple of pq_dim, a random rotation is always applied to the input data and queries to transform the working space from dim to rot_dim, which may be slightly larger than the original space and and is a multiple of pq_dim (rot_dim % pq_dim == 0). However, this transform is not necessary when dim is multiple of pq_dim (dim == rot_dim, hence no need in adding “extra” data columns / features).

By default, if dim == rot_dim, the rotation transform is initialized with the identity matrix. When force_random_rotation == true, a random orthogonal transform matrix is generated regardless of the values of dim and pq_dim.

bool conservative_memory_allocation = false#

By default, the algorithm allocates more space than necessary for individual clusters (list_data). This allows to amortize the cost of memory allocation and reduce the number of data copies during repeated calls to extend (extending the database).

The alternative is the conservative allocation behavior; when enabled, the algorithm always allocates the minimum amount of memory required to store the given number of records. Set this flag to true if you prefer to use as little GPU memory for the database as possible.

struct search_params : public raft::neighbors::ann::search_params#
#include <ivf_pq_types.hpp>

Public Members

uint32_t n_probes = 20#

The number of clusters to search.

cudaDataType_t lut_dtype = CUDA_R_32F#

Data type of look up table to be created dynamically at search time.

Possible values: [CUDA_R_32F, CUDA_R_16F, CUDA_R_8U]

The use of low-precision types reduces the amount of shared memory required at search time, so fast shared memory kernels can be used even for datasets with large dimansionality. Note that the recall is slightly degraded when low-precision type is selected.

cudaDataType_t internal_distance_dtype = CUDA_R_32F#

Storage data type for distance/similarity computed at search time.

Possible values: [CUDA_R_16F, CUDA_R_32F]

If the performance limiter at search time is device memory access, selecting FP16 will improve performance slightly.

double preferred_shmem_carveout = 1.0#

Preferred fraction of SM’s unified memory / L1 cache to be used as shared memory.

Possible values: [0.0 - 1.0] as a fraction of the sharedMemPerMultiprocessor.

One wants to increase the carveout to make sure a good GPU occupancy for the main search kernel, but not to keep it too high to leave some memory to be used as L1 cache. Note, this value is interpreted only as a hint. Moreover, a GPU usually allows only a fixed set of cache configurations, so the provided value is rounded up to the nearest configuration. Refer to the NVIDIA tuning guide for the target GPU architecture.

Note, this is a low-level tuning parameter that can have drastic negative effects on the search performance if tweaked incorrectly.

template<typename SizeT, typename IdxT>
struct list_spec#
#include <ivf_pq_types.hpp>

Public Types

using list_extents = extents<SizeT, dynamic_extent, dynamic_extent, kIndexGroupSize, kIndexGroupVecLen>#

PQ-encoded data stored in the interleaved format:

[ ceildiv(list_size, kIndexGroupSize) , ceildiv(pq_dim, (kIndexGroupVecLen * 8u) / pq_bits) , kIndexGroupSize , kIndexGroupVecLen ].

Public Functions

inline constexpr auto make_list_extents(SizeT n_rows) const -> list_extents#

Determine the extents of an array enough to hold a given amount of data.

template<typename IdxT>
struct index : public raft::neighbors::ann::index#
#include <ivf_pq_types.hpp>

IVF-PQ index.

In the IVF-PQ index, a database vector y is approximated with two level quantization:

y = Q_1(y) + Q_2(y - Q_1(y))

The first level quantizer (Q_1), maps the vector y to the nearest cluster center. The number of clusters is n_lists.

The second quantizer encodes the residual, and it is defined as a product quantizer [1].

A product quantizer encodes a dim dimensional vector with a pq_dim dimensional vector. First we split the input vector into pq_dim subvectors (denoted by u), where each u vector contains pq_len distinct components of y

y_1, y_2, … y_{pq_len}, y_{pq_len+1}, … y_{2*pq_len}, … y_{dim-pq_len+1} … y_{dim} ___________________/ ____________________________/ ______________________/ u_1 u_2 u_{pq_dim}

Then each subvector encoded with a separate quantizer q_i, end the results are concatenated

Q_2(y) = q_1(u_1),q_2(u_2),…,q_{pq_dim}(u_pq_dim})

Each quantizer q_i outputs a code with pq_bit bits. The second level quantizers are also defined by k-means clustering in the corresponding sub-space: the reproduction values are the centroids, and the set of reproduction values is the codebook.

When the data dimensionality dim is not multiple of pq_dim, the feature space is transformed using a random orthogonal matrix to have rot_dim = pq_dim * pq_len dimensions (rot_dim >= dim).

The second-level quantizers are trained either for each subspace or for each cluster: (a) codebook_gen::PER_SUBSPACE: creates pq_dim second-level quantizers - one for each slice of the data along features; (b) codebook_gen::PER_CLUSTER: creates n_lists second-level quantizers - one for each first-level cluster. In either case, the centroids are again found using k-means clustering interpreting the data as having pq_len dimensions.

[1] Product quantization for nearest neighbor search Herve Jegou, Matthijs Douze, Cordelia Schmid

Template Parameters:

IdxT – type of the indices in the source dataset

Public Functions

inline constexpr auto size() const noexcept -> IdxT#

Total length of the index.

inline constexpr auto dim() const noexcept -> uint32_t#

Dimensionality of the input data.

inline constexpr auto dim_ext() const noexcept -> uint32_t#

Dimensionality of the cluster centers: input data dim extended with vector norms and padded to 8 elems.

inline constexpr auto rot_dim() const noexcept -> uint32_t#

Dimensionality of the data after transforming it for PQ processing (rotated and augmented to be muplitple of pq_dim).

inline constexpr auto pq_bits() const noexcept -> uint32_t#

The bit length of an encoded vector element after compression by PQ.

inline constexpr auto pq_dim() const noexcept -> uint32_t#

The dimensionality of an encoded vector after compression by PQ.

inline constexpr auto pq_len() const noexcept -> uint32_t#

Dimensionality of a subspaces, i.e. the number of vector components mapped to a subspace

inline constexpr auto pq_book_size() const noexcept -> uint32_t#

The number of vectors in a PQ codebook (1 << pq_bits).

inline constexpr auto metric() const noexcept -> raft::distance::DistanceType#

Distance metric used for clustering.

inline constexpr auto codebook_kind() const noexcept -> codebook_gen#

How PQ codebooks are created.

inline constexpr auto n_lists() const noexcept -> uint32_t#

Number of clusters/inverted lists (first level quantization).

inline constexpr auto conservative_memory_allocation() const noexcept -> bool#

Whether to use convervative memory allocation when extending the list (cluster) data (see index_params.conservative_memory_allocation).

inline index(raft::resources const &handle, raft::distance::DistanceType metric, codebook_gen codebook_kind, uint32_t n_lists, uint32_t dim, uint32_t pq_bits = 8, uint32_t pq_dim = 0, bool conservative_memory_allocation = false)#

Construct an empty index. It needs to be trained and then populated.

inline index(raft::resources const &handle, const index_params &params, uint32_t dim)#

Construct an empty index. It needs to be trained and then populated.

inline auto pq_centers() noexcept -> device_mdspan<float, pq_centers_extents, row_major>#

PQ cluster centers

inline auto lists() noexcept -> std::vector<std::shared_ptr<list_data<IdxT>>>&#

Lists’ data and indices.

inline auto data_ptrs() noexcept -> device_vector_view<uint8_t*, uint32_t, row_major>#

Pointers to the inverted lists (clusters) data [n_lists].

inline auto inds_ptrs() noexcept -> device_vector_view<IdxT*, uint32_t, row_major>#

Pointers to the inverted lists (clusters) indices [n_lists].

inline auto rotation_matrix() noexcept -> device_matrix_view<float, uint32_t, row_major>#

The transform matrix (original space -> rotated padded space) [rot_dim, dim]

inline auto accum_sorted_sizes() noexcept -> host_vector_view<IdxT, uint32_t, row_major>#

Accumulated list sizes, sorted in descending order [n_lists + 1]. The last value contains the total length of the index. The value at index zero is always zero.

That is, the content of this span is as if the list_sizes was sorted and then accumulated.

This span is used during search to estimate the maximum size of the workspace.

inline auto list_sizes() noexcept -> device_vector_view<uint32_t, uint32_t, row_major>#

Sizes of the lists [n_lists].

inline auto centers() noexcept -> device_matrix_view<float, uint32_t, row_major>#

Cluster centers corresponding to the lists in the original space [n_lists, dim_ext]

inline auto centers_rot() noexcept -> device_matrix_view<float, uint32_t, row_major>#

Cluster centers corresponding to the lists in the rotated space [n_lists, rot_dim]

inline auto get_list_size_in_bytes(uint32_t label) -> uint32_t#

fetch size of a particular IVF list in bytes using the list extents. Usage example:

raft::resources res;
// use default index params
ivf_pq::index_params index_params;
// extend the IVF lists while building the index
index_params.add_data_on_build = true;
// create and fill the index from a [N, D] dataset
auto index = raft::neighbors::ivf_pq::build<int64_t>(res, index_params, dataset, N, D);
// Fetch the size of the fourth list
uint32_t size = index.get_list_size_in_bytes(3);

Parameters:

label[in] list ID

Serializer Methods#

#include <raft/neighbors/ivf_pq_serialize.cuh>

namespace raft::neighbors::ivf_pq

template<typename IdxT>
void serialize(raft::resources const &handle, std::ostream &os, const index<IdxT> &index)#

Write the index to an output stream

Experimental, both the API and the serialization format are subject to change.

#include <raft/core/resources.hpp>

raft::resources handle;

// create an output stream
std::ostream os(std::cout.rdbuf());
// create an index with `auto index = ivf_pq::build(...);`
raft::serialize(handle, os, index);
Template Parameters:

IdxT – type of the index

Parameters:
  • handle[in] the raft handle

  • os[in] output stream

  • index[in] IVF-PQ index

template<typename IdxT>
void serialize(raft::resources const &handle, const std::string &filename, const index<IdxT> &index)#

Save the index to file.

Experimental, both the API and the serialization format are subject to change.

#include <raft/core/resources.hpp>

raft::resources handle;

// create a string with a filepath
std::string filename("/path/to/index");
// create an index with `auto index = ivf_pq::build(...);`
raft::serialize(handle, filename, index);
Template Parameters:

IdxT – type of the index

Parameters:
  • handle[in] the raft handle

  • filename[in] the file name for saving the index

  • index[in] IVF-PQ index

template<typename IdxT>
index<IdxT> deserialize(raft::resources const &handle, std::istream &is)#

Load index from input stream

Experimental, both the API and the serialization format are subject to change.

#include <raft/core/resources.hpp>

raft::resources handle;

// create an input stream
std::istream is(std::cin.rdbuf());
using IdxT = int; // type of the index
auto index = raft::deserialize<IdxT>(handle, is);
Template Parameters:

IdxT – type of the index

Parameters:
  • handle[in] the raft handle

  • is[in] input stream

Returns:

raft::neighbors::ivf_pq::index<IdxT>

template<typename IdxT>
index<IdxT> deserialize(raft::resources const &handle, const std::string &filename)#

Load index from file.

Experimental, both the API and the serialization format are subject to change.

#include <raft/core/resources.hpp>

raft::resources handle;

// create a string with a filepath
std::string filename("/path/to/index");
using IdxT = int; // type of the index
auto index = raft::deserialize<IdxT>(handle, filename);
Template Parameters:

IdxT – type of the index

Parameters:
  • handle[in] the raft handle

  • filename[in] the name of the file that stores the index

Returns:

raft::neighbors::ivf_pq::index<IdxT>

Candidate Refinement#

#include <raft/neighbors/refine.cuh>

namespace raft::neighbors

template<typename idx_t, typename data_t, typename distance_t, typename matrix_idx>
void refine(raft::resources const &handle, raft::device_matrix_view<const data_t, matrix_idx, row_major> dataset, raft::device_matrix_view<const data_t, matrix_idx, row_major> queries, raft::device_matrix_view<const idx_t, matrix_idx, row_major> neighbor_candidates, raft::device_matrix_view<idx_t, matrix_idx, row_major> indices, raft::device_matrix_view<distance_t, matrix_idx, row_major> distances, distance::DistanceType metric = distance::DistanceType::L2Unexpanded)#

Refine nearest neighbor search.

Refinement is an operation that follows an approximate NN search. The approximate search has already selected n_candidates neighbor candidates for each query. We narrow it down to k neighbors. For each query, we calculate the exact distance between the query and its n_candidates neighbor candidate, and select the k nearest ones.

The k nearest neighbors and distances are returned.

Example usage

using namespace raft::neighbors;
// use default index parameters
ivf_pq::index_params index_params;
// create and fill the index from a [N, D] dataset
auto index = ivf_pq::build(handle, index_params, dataset, N, D);
// use default search parameters
ivf_pq::search_params search_params;
// search m = 4 * k nearest neighbours for each of the N queries
ivf_pq::search(handle, search_params, index, queries, N, 4 * k, neighbor_candidates,
               out_dists_tmp);
// refine it to the k nearest one
refine(handle, dataset, queries, neighbor_candidates, out_indices, out_dists,
        index.metric());

Parameters:
  • handle[in] the raft handle

  • dataset[in] device matrix that stores the dataset [n_rows, dims]

  • queries[in] device matrix of the queries [n_queris, dims]

  • neighbor_candidates[in] indices of candidate vectors [n_queries, n_candidates], where n_candidates >= k

  • indices[out] device matrix that stores the refined indices [n_queries, k]

  • distances[out] device matrix that stores the refined distances [n_queries, k]

  • metric[in] distance metric to use. Euclidean (L2) is used by default

template<typename idx_t, typename data_t, typename distance_t, typename matrix_idx>
void refine(raft::resources const &handle, raft::host_matrix_view<const data_t, matrix_idx, row_major> dataset, raft::host_matrix_view<const data_t, matrix_idx, row_major> queries, raft::host_matrix_view<const idx_t, matrix_idx, row_major> neighbor_candidates, raft::host_matrix_view<idx_t, matrix_idx, row_major> indices, raft::host_matrix_view<distance_t, matrix_idx, row_major> distances, distance::DistanceType metric = distance::DistanceType::L2Unexpanded)#

Same as above, but all input and out data is in host memory.

Parameters:
  • handle[in] the raft handle

  • dataset[in] host matrix that stores the dataset [n_rows, dims]

  • queries[in] host matrix of the queries [n_queris, dims]

  • neighbor_candidates[in] host matrix with indices of candidate vectors [n_queries, n_candidates], where n_candidates >= k

  • indices[out] host matrix that stores the refined indices [n_queries, k]

  • distances[out] host matrix that stores the refined distances [n_queries, k]

  • metric[in] distance metric to use. Euclidean (L2) is used by default

Helper Methods#

#include <raft/neighbors/ivf_pq_helpers.cuh>

namespace raft::neighbors::ivf_pq::helpers

template<typename IdxT>
void pack_list_data(raft::resources const &res, index<IdxT> *index, device_matrix_view<const uint8_t, uint32_t, row_major> codes, uint32_t label, uint32_t offset)#

Write flat PQ codes into an existing list by the given offset.

The list is identified by its label.

NB: no memory allocation happens here; the list must fit the data (offset + n_vec).

Usage example:

// We will write into the 137th cluster
uint32_t label = 137;
// allocate the buffer for the input codes
auto codes = raft::make_device_matrix<const uint8_t>(res, n_vec, index.pq_dim());
... prepare n_vecs to pack into the list in codes ...
// write codes into the list starting from the 42nd position
ivf_pq::helpers::pack_list_data(res, &index, codes_to_pack, label, 42);

Parameters:
  • res[in] raft resource

  • index[inout] IVF-PQ index.

  • codes[in] flat PQ codes, one code per byte [n_rows, pq_dim]

  • label[in] The id of the list (cluster) into which we write.

  • offset[in] how many records to skip before writing the data into the list

template<typename IdxT>
void pack_contiguous_list_data(raft::resources const &res, index<IdxT> *index, uint8_t *codes, uint32_t n_rows, uint32_t label, uint32_t offset)#

Write flat PQ codes into an existing list by the given offset. Use this when the input vectors are PQ encoded and not expanded to one code per byte.

The list is identified by its label.

NB: no memory allocation happens here; the list into which the vectors are packed must fit offset

  • n_rows rows.

Usage example:

using namespace raft::neighbors;
raft::resources res;
// use default index parameters
ivf_pq::index_params index_params;
// create and fill the index from a [N, D] dataset
auto index = ivf_pq::build(res, index_params, dataset, N, D);
// allocate the buffer for n_rows input codes. Each vector occupies
// raft::ceildiv(index.pq_dim() * index.pq_bits(), 8) bytes because
// codes are compressed and without gaps.
auto codes = raft::make_device_matrix<const uint8_t>(
  res, n_rows, raft::ceildiv(index.pq_dim() * index.pq_bits(), 8));
... prepare the compressed vectors to pack into the list in codes ...
// the first n_rows codes in the fourth IVF list are to be overwritten.
uint32_t label = 3;
// write codes into the list starting from the 0th position
ivf_pq::helpers::pack_contiguous_list_data(
  res, &index, codes.data_handle(), n_rows, label, 0);

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[inout] pointer to IVF-PQ index

  • codes[in] flat contiguous PQ codes [n_rows, ceildiv(pq_dim * pq_bits, 8)]

  • n_rows[in] how many records to pack

  • label[in] The id of the list (cluster) into which we write.

  • offset[in] how many records to skip before writing the data into the list

template<typename IdxT>
void unpack_list_data(raft::resources const &res, const index<IdxT> &index, device_matrix_view<uint8_t, uint32_t, row_major> out_codes, uint32_t label, uint32_t offset)#

Unpack n_take consecutive records of a single list (cluster) in the compressed index starting at given offset, one code per byte (independently of pq_bits).

Usage example:

  // We will unpack the fourth cluster
  uint32_t label = 3;
  // Get the list size
  uint32_t list_size = 0;
  raft::copy(&list_size, index.list_sizes().data_handle() + label, 1,
resource::get_cuda_stream(res)); resource::sync_stream(res);
  // allocate the buffer for the output
  auto codes = raft::make_device_matrix<float>(res, list_size, index.pq_dim());
  // unpack the whole list
  ivf_pq::helpers::unpack_list_data(res, index, codes.view(), label, 0);

Template Parameters:

IdxT – type of the indices in the source dataset

Parameters:
  • res[in]

  • index[in]

  • out_codes[out] the destination buffer [n_take, index.pq_dim()]. The length n_take defines how many records to unpack, it must be smaller than the list size.

  • label[in] The id of the list (cluster) to decode.

  • offset[in] How many records in the list to skip.

template<typename IdxT>
void unpack_list_data(raft::resources const &res, const index<IdxT> &index, device_vector_view<const uint32_t> in_cluster_indices, device_matrix_view<uint8_t, uint32_t, row_major> out_codes, uint32_t label)#

Unpack a series of records of a single list (cluster) in the compressed index by their in-list offsets, one code per byte (independently of pq_bits).

Usage example:

// We will unpack the fourth cluster
uint32_t label = 3;
// Create the selection vector
auto selected_indices = raft::make_device_vector<uint32_t>(res, 4);
... fill the indices ...
resource::sync_stream(res);
// allocate the buffer for the output
auto codes = raft::make_device_matrix<float>(res, selected_indices.size(), index.pq_dim());
// decode the whole list
ivf_pq::helpers::unpack_list_data(
    res, index, selected_indices.view(), codes.view(), label);

Template Parameters:

IdxT – type of the indices in the source dataset

Parameters:
  • res[in] raft resource

  • index[in] IVF-PQ index (passed by reference)

  • in_cluster_indices[in] The offsets of the selected indices within the cluster.

  • out_codes[out] the destination buffer [n_take, index.pq_dim()]. The length n_take defines how many records to unpack, it must be smaller than the list size.

  • label[in] The id of the list (cluster) to decode.

template<typename IdxT>
void unpack_contiguous_list_data(raft::resources const &res, const index<IdxT> &index, uint8_t *out_codes, uint32_t n_rows, uint32_t label, uint32_t offset)#

Unpack n_rows consecutive PQ encoded vectors of a single list (cluster) in the compressed index starting at given offset, not expanded to one code per byte. Each code in the output buffer occupies ceildiv(index.pq_dim() * index.pq_bits(), 8) bytes.

Usage example:

  raft::resources res;
  // We will unpack the whole fourth cluster
  uint32_t label = 3;
  // Get the list size
  uint32_t list_size = 0;
  raft::update_host(&list_size, index.list_sizes().data_handle() + label, 1,
raft::resource::get_cuda_stream(res)); raft::resource::sync_stream(res);
  // allocate the buffer for the output
  auto codes = raft::make_device_matrix<float>(res, list_size, raft::ceildiv(index.pq_dim() *
index.pq_bits(), 8));
  // unpack the whole list
  ivf_pq::helpers::unpack_list_data(res, index, codes.data_handle(), list_size, label, 0);

Template Parameters:

IdxT – type of the indices in the source dataset

Parameters:
  • res[in] raft resource

  • index[in] IVF-PQ index (passed by reference)

  • out_codes[out] the destination buffer [n_rows, ceildiv(index.pq_dim() * index.pq_bits(), 8)]. The length n_rows defines how many records to unpack, offset + n_rows must be smaller than or equal to the list size.

  • n_rows[in] how many codes to unpack

  • label[in] The id of the list (cluster) to decode.

  • offset[in] How many records in the list to skip.

template<typename T, typename IdxT>
void reconstruct_list_data(raft::resources const &res, const index<IdxT> &index, device_matrix_view<T, uint32_t, row_major> out_vectors, uint32_t label, uint32_t offset)#

Decode n_take consecutive records of a single list (cluster) in the compressed index starting at given offset.

Usage example:

// We will reconstruct the fourth cluster
uint32_t label = 3;
// Get the list size
uint32_t list_size = 0;
raft::copy(&list_size, index.list_sizes().data_handle() + label, 1,
resource::get_cuda_stream(res)); resource::sync_stream(res);
// allocate the buffer for the output
auto decoded_vectors = raft::make_device_matrix<float>(res, list_size, index.dim());
// decode the whole list
ivf_pq::helpers::reconstruct_list_data(res, index, decoded_vectors.view(), label, 0);

Template Parameters:
  • T – data element type

  • IdxT – type of the indices in the source dataset

Parameters:
  • res[in]

  • index[in]

  • out_vectors[out] the destination buffer [n_take, index.dim()]. The length n_take defines how many records to reconstruct, it must be smaller than the list size.

  • label[in] The id of the list (cluster) to decode.

  • offset[in] How many records in the list to skip.

template<typename T, typename IdxT>
void reconstruct_list_data(raft::resources const &res, const index<IdxT> &index, device_vector_view<const uint32_t> in_cluster_indices, device_matrix_view<T, uint32_t, row_major> out_vectors, uint32_t label)#

Decode a series of records of a single list (cluster) in the compressed index by their in-list offsets.

Usage example:

// We will reconstruct the fourth cluster
uint32_t label = 3;
// Create the selection vector
auto selected_indices = raft::make_device_vector<uint32_t>(res, 4);
... fill the indices ...
resource::sync_stream(res);
// allocate the buffer for the output
auto decoded_vectors = raft::make_device_matrix<float>(
                          res, selected_indices.size(), index.dim());
// decode the whole list
ivf_pq::helpers::reconstruct_list_data(
    res, index, selected_indices.view(), decoded_vectors.view(), label);

Template Parameters:
  • T – data element type

  • IdxT – type of the indices in the source dataset

Parameters:
  • res[in]

  • index[in]

  • in_cluster_indices[in] The offsets of the selected indices within the cluster.

  • out_vectors[out] the destination buffer [n_take, index.dim()]. The length n_take defines how many records to reconstruct, it must be smaller than the list size.

  • label[in] The id of the list (cluster) to decode.

template<typename IdxT>
void extend_list_with_codes(raft::resources const &res, index<IdxT> *index, device_matrix_view<const uint8_t, uint32_t, row_major> new_codes, device_vector_view<const IdxT, uint32_t, row_major> new_indices, uint32_t label)#

Extend one list of the index in-place, by the list label, skipping the classification and encoding steps.

Usage example:

// We will extend the fourth cluster
uint32_t label = 3;
// We will fill 4 new vectors
uint32_t n_vec = 4;
// Indices of the new vectors
auto indices = raft::make_device_vector<uint32_t>(res, n_vec);
... fill the indices ...
auto new_codes = raft::make_device_matrix<uint8_t, uint32_t, row_major> new_codes(
    res, n_vec, index.pq_dim());
... fill codes ...
// extend list with new codes
ivf_pq::helpers::extend_list_with_codes(
    res, &index, codes.view(), indices.view(), label);

Template Parameters:

IdxT

Parameters:
  • res[in]

  • index[inout]

  • new_codes[in] flat PQ codes, one code per byte [n_rows, index.pq_dim()]

  • new_indices[in] source indices [n_rows]

  • label[in] the id of the target list (cluster).

template<typename T, typename IdxT>
void extend_list(raft::resources const &res, index<IdxT> *index, device_matrix_view<const T, uint32_t, row_major> new_vectors, device_vector_view<const IdxT, uint32_t, row_major> new_indices, uint32_t label)#

Extend one list of the index in-place, by the list label, skipping the classification step.

Usage example:

// We will extend the fourth cluster
uint32_t label = 3;
// We will extend with 4 new vectors
uint32_t n_vec = 4;
// Indices of the new vectors
auto indices = raft::make_device_vector<uint32_t>(res, n_vec);
... fill the indices ...
auto new_vectors = raft::make_device_matrix<float, uint32_t, row_major> new_codes(
    res, n_vec, index.dim());
... fill vectors ...
// extend list with new vectors
ivf_pq::helpers::extend_list(
    res, &index, new_vectors.view(), indices.view(), label);

Template Parameters:
  • T

  • IdxT

Parameters:
  • res[in]

  • index[inout]

  • new_vectors[in] data to encode [n_rows, index.dim()]

  • new_indices[in] source indices [n_rows]

  • label[in] the id of the target list (cluster).

template<typename IdxT>
void erase_list(raft::resources const &res, index<IdxT> *index, uint32_t label)#

Remove all data from a single list (cluster) in the index.

Usage example:

// We will erase the fourth cluster (label = 3)
ivf_pq::helpers::erase_list(res, &index, 3);

Template Parameters:

IdxT

Parameters:
  • res[in]

  • index[inout]

  • label[in] the id of the target list (cluster).

template<typename IdxT>
void reset_index(const raft::resources &res, index<IdxT> *index)#

Public helper API to reset the data and indices ptrs, and the list sizes. Useful for externally modifying the index without going through the build stage. The data and indices of the IVF lists will be lost.

Usage example:

raft::resources res;
using namespace raft::neighbors;
// use default index parameters
ivf_pq::index_params index_params;
// initialize an empty index
ivf_pq::index<int64_t> index(res, index_params, D);
// reset the index's state and list sizes
ivf_pq::helpers::reset_index(res, &index);

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[inout] pointer to IVF-PQ index

template<typename IdxT>
void make_rotation_matrix(raft::resources const &res, index<IdxT> *index, bool force_random_rotation)#

Public helper API exposing the computation of the index’s rotation matrix. NB: This is to be used only when the rotation matrix is not already computed through raft::neighbors::ivf_pq::build.

Usage example:

raft::resources res;
// use default index parameters
ivf_pq::index_params index_params;
// force random rotation
index_params.force_random_rotation = true;
// initialize an empty index
raft::neighbors::ivf_pq::index<int64_t> index(res, index_params, D);
// reset the index
reset_index(res, &index);
// compute the rotation matrix with random_rotation
raft::neighbors::ivf_pq::helpers::make_rotation_matrix(
  res, &index, index_params.force_random_rotation);

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[inout] pointer to IVF-PQ index

  • force_random_rotation[in] whether to apply a random rotation matrix on the input data. See raft::neighbors::ivf_pq::index_params for more details.

template<typename IdxT>
void set_centers(raft::resources const &res, index<IdxT> *index, device_matrix_view<const float, uint32_t> cluster_centers)#

Public helper API for externally modifying the index’s IVF centroids. NB: The index must be reset before this. Use raft::neighbors::ivf_pq::extend to construct IVF lists according to new centroids.

Usage example:

    raft::resources res;
    // allocate the buffer for the input centers
    auto cluster_centers = raft::make_device_matrix<float, uint32_t>(res, index.n_lists(),
index.dim());
    ... prepare ivf centroids in cluster_centers ...
    // reset the index
    reset_index(res, &index);
    // recompute the state of the index
    raft::neighbors::ivf_pq::helpers::recompute_internal_state(res, index);
    // Write the IVF centroids
    raft::neighbors::ivf_pq::helpers::set_centers(
                   res,
                   &index,
                   cluster_centers);

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[inout] pointer to IVF-PQ index

  • cluster_centers[in] new cluster centers [index.n_lists(), index.dim()]

template<typename IdxT>
void recompute_internal_state(const raft::resources &res, index<IdxT> *index)#

Helper exposing the re-computation of list sizes and related arrays if IVF lists have been modified.

Usage example:

using namespace raft::neighbors;
raft::resources res;
// use default index parameters
ivf_pq::index_params index_params;
// initialize an empty index
ivf_pq::index<int64_t> index(res, index_params, D);
ivf_pq::helpers::reset_index(res, &index);
// resize the first IVF list to hold 5 records
auto spec = list_spec<uint32_t, int64_t>{
  index->pq_bits(), index->pq_dim(), index->conservative_memory_allocation()};
uint32_t new_size = 5;
ivf::resize_list(res, list, spec, new_size, 0);
raft::update_device(index.list_sizes(), &new_size, 1, stream);
// recompute the internal state of the index
ivf_pq::helpers::recompute_internal_state(res, &index);

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[inout] pointer to IVF-PQ index

template<typename IdxT>
void extract_centers(raft::resources const &res, const index<IdxT> &index, raft::device_matrix_view<float> cluster_centers)#

Public helper API for fetching a trained index’s IVF centroids into a buffer that may be allocated on either host or device.

Usage example:

raft::resources res;
// allocate the buffer for the output centers
auto cluster_centers = raft::make_device_matrix<float, uint32_t>(
  res, index.n_lists(), index.dim());
// Extract the IVF centroids into the buffer
raft::neighbors::ivf_pq::helpers::extract_centers(res, index, cluster_centers.data_handle());

Template Parameters:

IdxT

Parameters:
  • res[in] raft resource

  • index[in] IVF-PQ index (passed by reference)

  • cluster_centers[out] IVF cluster centers [index.n_lists(), index.dim]